I have been working with Flux LoRAs for not very long, but have been on and off focused on creating a general male realistic Flux Dev-1 model.

Over 4k images trained between two LoRAs (a merge for now, until I retrain) over 60k steps. A wide variety from Reddit (before the API got locked down) as well as several site scrapes for male photography. I ran everything through JoyCaption and hand checked over 1k captions for accuracy.

This is my a merge of my earlier work focusing on vintage male pornography and modern male selfie pornography... The result... something that can make penises with reliable accuracy but with a limited range of known sexual acts.

This model is geared more towards male nudity in general, and not lewd acts with other men (although the dataset did contain a good amount of male-on-male action).

This is the first public release for any of my NSFW models and I'd appreciate honest feedback. If its not too your tastes, no biggie (a lot of other work is definitely not in my wheelhouse). I ask that if you do use the model, please provide valid critism so that I may address it.

Weakness:

Strengths:

This is not a "specific" model. It does not do any one thing well. Please do not expect it to make some niche fetish pornography (although if it does, I'd be willing to make a fork of it pushing that direction)

Consider this model MIT, in that you may modify, merge, and redistribute as long as you give me a tiny shoutout.

Trigger words (because its a merge for now): `fknxtm` and `fkosm`

The trigger word `fkosm` is more important for correct dick generations, this is because `fkosm` was a more distinct dataset for vintage dicks, and `fknxtm` is more of a general set of men in a variety of poses which caused the training to not be as precise.

Planning to retrain on RunPod for a longer training time and more epochs, the current version has been trained on Google Colab for approx 24 hours (max time).

I'd start at weight 0.8 and go from there. If you get limp dicks, try a higher weight, if you get dick fingers, go down.

I'd like to think the men generated are vastly different from base model plastiques, but you tell me.

Don't count toes (how strange with Flux)

If you like vintage photos use only trigger word `fkosm` if you like modern males, use `fknxtm, fkosm` (fksosm keep dicks in dick shape; retraining on trigger fknxtm)

You can try combining both trigger words with just `fk`, though I'd suggest using both, `fknxtm, fkosm, a photo of a nude man...`

I plan on training a custom version of JoyCaption using the available training script to more closely match the male anatomy and common poses. From testing "dicks peeking through shorts" turn to "shorts pulled down revealing penis" (which is how you should caption this model--currently) as well as tons of other male anatomy pose quirks. If you'd like to contribute to training a custom JoyCaption, please human label all of your images, and send them to me in a PM.

``` Full body shot, powerful man with a broad chest and a stoic expression, standing under a cold, driving rain. His nude body is glistening, with every muscle defined by the downpour, his thick, semi-erect penis visibly adhering to his leg, a raw and undeniable display of his resilience. Background is a stormy, windswept city street at night, with blurred lights and reflections, conveying raw power and resilience, powerfully mirroring his unwavering and exposed strength. ```

You should use fkosm to trigger the image generation.

You should use fknxtm to trigger the image generation.

Weights for this model are available in Safetensors format.

Download them in the Files & versions tab.